Video concept classification and retrieval

Search and retrieval in video content is a trending topic in computer vision. Difficulties of this research topic is two folds; extracting semantic information from structure of video images is not a simple task and demanding nature of video content requires efficient algorithms. Semantic information extraction is challenged by researchers for more than two decades, yet new improvements are still welcome by the community. Recent burst of efficient computer hardware architectures has exploited both accuracy and complexity of many algorithms adding a new dimension to the efficient algorithm selection. In this thesis, our goal is to classify visual concepts in video data for content-based search and retrieval applications. To this end, we introduce a complete visual concept classification and retrieval system. We use two state-of-the-art methods, namely "Bag-of-Words" (BoW) and "Convolutional Neural Network" (CNN) architecture for visual concept classification. The performance of the classifiers is further improved by optimizing the processing pipeline steps. For retrieval, we provide concept- and content- based querying of video data and perform evaluations on Oxford Buildings and Paris datasets. Results show that, a substantial performance gain is possible by optimizing processing pipelines of the classifiers and deep learning based methods outperform the BoW.